Expand MLA to support most types of quantization (#13181)

vllm/config.py

@@ -991,37 +991,7 @@ class ModelConfig:
 
     @property
     def use_mla(self) -> bool:
-        if not self.is_deepseek_mla or envs.VLLM_MLA_DISABLE:
-            return False
-
-        if self.quantization is not None and self.quantization not in [\
-            "fp8", "compressed-tensors"]:
-            logger.warning(
-                "MLA is not supported with %s quantization. "
-                "Disabling MLA.", self.quantization)
-            return False
-
-        # If using a "compressed-tensors" checkpoint, check that all groups
-        # have fp8 for both weights and activations.
-        if self.quantization == "compressed-tensors":
-            quant_config = self._parse_quant_hf_config()
-            for group_name, cfg in quant_config.get("config_groups", {
-                    "": {}
-            }).items():
-                act_cfg = cfg.get("input_activations", {})
-                act_type = None if act_cfg is None else act_cfg.get("type", "")
-                w_cfg = cfg.get("weights", {})
-                w_type = None if w_cfg is None else w_cfg.get("type", "")
-                if act_type != "fp8" or w_type != "fp8":
-                    logger.warning(
-                        "compressed-tensors MLA support requires fp8 "
-                        "activations and weights in group '%s', but got "
-                        "activations type '%s' and weights type '%s'.\n "
-                        "Full config: %s", group_name, act_type, w_type,
-                        quant_config)
-                    return False
-
-        return True
+        return self.is_deepseek_mla and not envs.VLLM_MLA_DISABLE
 
     @property
     def supported_runner_types(self) -> Set[RunnerType]: